Commit | Line | Data |
---|---|---|
e43473b7 VG |
1 | /* |
2 | * Interface for controlling IO bandwidth on a request queue | |
3 | * | |
4 | * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com> | |
5 | */ | |
6 | ||
7 | #include <linux/module.h> | |
8 | #include <linux/slab.h> | |
9 | #include <linux/blkdev.h> | |
10 | #include <linux/bio.h> | |
11 | #include <linux/blktrace_api.h> | |
12 | #include "blk-cgroup.h" | |
bc9fcbf9 | 13 | #include "blk.h" |
e43473b7 VG |
14 | |
15 | /* Max dispatch from a group in 1 round */ | |
16 | static int throtl_grp_quantum = 8; | |
17 | ||
18 | /* Total max dispatch from all groups in one round */ | |
19 | static int throtl_quantum = 32; | |
20 | ||
21 | /* Throttling is performed over 100ms slice and after that slice is renewed */ | |
22 | static unsigned long throtl_slice = HZ/10; /* 100 ms */ | |
23 | ||
3c798398 | 24 | static struct blkcg_policy blkcg_policy_throtl; |
0381411e | 25 | |
450adcbe VG |
26 | /* A workqueue to queue throttle related work */ |
27 | static struct workqueue_struct *kthrotld_workqueue; | |
450adcbe | 28 | |
c9e0332e TH |
29 | struct throtl_service_queue { |
30 | struct rb_root pending_tree; /* RB tree of active tgs */ | |
31 | struct rb_node *first_pending; /* first node in the tree */ | |
32 | unsigned int nr_pending; /* # queued in the tree */ | |
33 | unsigned long first_pending_disptime; /* disptime of the first tg */ | |
e43473b7 VG |
34 | }; |
35 | ||
c9e0332e TH |
36 | #define THROTL_SERVICE_QUEUE_INITIALIZER \ |
37 | (struct throtl_service_queue){ .pending_tree = RB_ROOT } | |
e43473b7 | 38 | |
5b2c16aa TH |
39 | enum tg_state_flags { |
40 | THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */ | |
41 | }; | |
42 | ||
e43473b7 VG |
43 | #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node) |
44 | ||
8a3d2615 TH |
45 | /* Per-cpu group stats */ |
46 | struct tg_stats_cpu { | |
47 | /* total bytes transferred */ | |
48 | struct blkg_rwstat service_bytes; | |
49 | /* total IOs serviced, post merge */ | |
50 | struct blkg_rwstat serviced; | |
51 | }; | |
52 | ||
e43473b7 | 53 | struct throtl_grp { |
f95a04af TH |
54 | /* must be the first member */ |
55 | struct blkg_policy_data pd; | |
56 | ||
c9e0332e | 57 | /* active throtl group service_queue member */ |
e43473b7 VG |
58 | struct rb_node rb_node; |
59 | ||
60 | /* | |
61 | * Dispatch time in jiffies. This is the estimated time when group | |
62 | * will unthrottle and is ready to dispatch more bio. It is used as | |
63 | * key to sort active groups in service tree. | |
64 | */ | |
65 | unsigned long disptime; | |
66 | ||
e43473b7 VG |
67 | unsigned int flags; |
68 | ||
69 | /* Two lists for READ and WRITE */ | |
70 | struct bio_list bio_lists[2]; | |
71 | ||
72 | /* Number of queued bios on READ and WRITE lists */ | |
73 | unsigned int nr_queued[2]; | |
74 | ||
75 | /* bytes per second rate limits */ | |
76 | uint64_t bps[2]; | |
77 | ||
8e89d13f VG |
78 | /* IOPS limits */ |
79 | unsigned int iops[2]; | |
80 | ||
e43473b7 VG |
81 | /* Number of bytes disptached in current slice */ |
82 | uint64_t bytes_disp[2]; | |
8e89d13f VG |
83 | /* Number of bio's dispatched in current slice */ |
84 | unsigned int io_disp[2]; | |
e43473b7 VG |
85 | |
86 | /* When did we start a new slice */ | |
87 | unsigned long slice_start[2]; | |
88 | unsigned long slice_end[2]; | |
fe071437 | 89 | |
8a3d2615 TH |
90 | /* Per cpu stats pointer */ |
91 | struct tg_stats_cpu __percpu *stats_cpu; | |
92 | ||
93 | /* List of tgs waiting for per cpu stats memory to be allocated */ | |
94 | struct list_head stats_alloc_node; | |
e43473b7 VG |
95 | }; |
96 | ||
97 | struct throtl_data | |
98 | { | |
e43473b7 | 99 | /* service tree for active throtl groups */ |
c9e0332e | 100 | struct throtl_service_queue service_queue; |
e43473b7 | 101 | |
e43473b7 VG |
102 | struct request_queue *queue; |
103 | ||
104 | /* Total Number of queued bios on READ and WRITE lists */ | |
105 | unsigned int nr_queued[2]; | |
106 | ||
107 | /* | |
02977e4a | 108 | * number of total undestroyed groups |
e43473b7 VG |
109 | */ |
110 | unsigned int nr_undestroyed_grps; | |
111 | ||
112 | /* Work for dispatching throttled bios */ | |
cb76199c | 113 | struct delayed_work dispatch_work; |
e43473b7 VG |
114 | }; |
115 | ||
8a3d2615 TH |
116 | /* list and work item to allocate percpu group stats */ |
117 | static DEFINE_SPINLOCK(tg_stats_alloc_lock); | |
118 | static LIST_HEAD(tg_stats_alloc_list); | |
119 | ||
120 | static void tg_stats_alloc_fn(struct work_struct *); | |
121 | static DECLARE_DELAYED_WORK(tg_stats_alloc_work, tg_stats_alloc_fn); | |
122 | ||
f95a04af TH |
123 | static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd) |
124 | { | |
125 | return pd ? container_of(pd, struct throtl_grp, pd) : NULL; | |
126 | } | |
127 | ||
3c798398 | 128 | static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg) |
0381411e | 129 | { |
f95a04af | 130 | return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl)); |
0381411e TH |
131 | } |
132 | ||
3c798398 | 133 | static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg) |
0381411e | 134 | { |
f95a04af | 135 | return pd_to_blkg(&tg->pd); |
0381411e TH |
136 | } |
137 | ||
03d8e111 TH |
138 | static inline struct throtl_grp *td_root_tg(struct throtl_data *td) |
139 | { | |
140 | return blkg_to_tg(td->queue->root_blkg); | |
141 | } | |
142 | ||
54e7ed12 TH |
143 | #define throtl_log_tg(td, tg, fmt, args...) do { \ |
144 | char __pbuf[128]; \ | |
145 | \ | |
146 | blkg_path(tg_to_blkg(tg), __pbuf, sizeof(__pbuf)); \ | |
147 | blk_add_trace_msg((td)->queue, "throtl %s " fmt, __pbuf, ##args); \ | |
148 | } while (0) | |
e43473b7 VG |
149 | |
150 | #define throtl_log(td, fmt, args...) \ | |
151 | blk_add_trace_msg((td)->queue, "throtl " fmt, ##args) | |
152 | ||
8a3d2615 TH |
153 | /* |
154 | * Worker for allocating per cpu stat for tgs. This is scheduled on the | |
3b07e9ca | 155 | * system_wq once there are some groups on the alloc_list waiting for |
8a3d2615 TH |
156 | * allocation. |
157 | */ | |
158 | static void tg_stats_alloc_fn(struct work_struct *work) | |
159 | { | |
160 | static struct tg_stats_cpu *stats_cpu; /* this fn is non-reentrant */ | |
161 | struct delayed_work *dwork = to_delayed_work(work); | |
162 | bool empty = false; | |
163 | ||
164 | alloc_stats: | |
165 | if (!stats_cpu) { | |
166 | stats_cpu = alloc_percpu(struct tg_stats_cpu); | |
167 | if (!stats_cpu) { | |
168 | /* allocation failed, try again after some time */ | |
3b07e9ca | 169 | schedule_delayed_work(dwork, msecs_to_jiffies(10)); |
8a3d2615 TH |
170 | return; |
171 | } | |
172 | } | |
173 | ||
174 | spin_lock_irq(&tg_stats_alloc_lock); | |
175 | ||
176 | if (!list_empty(&tg_stats_alloc_list)) { | |
177 | struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list, | |
178 | struct throtl_grp, | |
179 | stats_alloc_node); | |
180 | swap(tg->stats_cpu, stats_cpu); | |
181 | list_del_init(&tg->stats_alloc_node); | |
182 | } | |
183 | ||
184 | empty = list_empty(&tg_stats_alloc_list); | |
185 | spin_unlock_irq(&tg_stats_alloc_lock); | |
186 | if (!empty) | |
187 | goto alloc_stats; | |
188 | } | |
189 | ||
3c798398 | 190 | static void throtl_pd_init(struct blkcg_gq *blkg) |
a29a171e | 191 | { |
0381411e | 192 | struct throtl_grp *tg = blkg_to_tg(blkg); |
ff26eaad | 193 | unsigned long flags; |
cd1604fa | 194 | |
a29a171e VG |
195 | RB_CLEAR_NODE(&tg->rb_node); |
196 | bio_list_init(&tg->bio_lists[0]); | |
197 | bio_list_init(&tg->bio_lists[1]); | |
a29a171e | 198 | |
e56da7e2 TH |
199 | tg->bps[READ] = -1; |
200 | tg->bps[WRITE] = -1; | |
201 | tg->iops[READ] = -1; | |
202 | tg->iops[WRITE] = -1; | |
8a3d2615 TH |
203 | |
204 | /* | |
205 | * Ugh... We need to perform per-cpu allocation for tg->stats_cpu | |
206 | * but percpu allocator can't be called from IO path. Queue tg on | |
207 | * tg_stats_alloc_list and allocate from work item. | |
208 | */ | |
ff26eaad | 209 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); |
8a3d2615 | 210 | list_add(&tg->stats_alloc_node, &tg_stats_alloc_list); |
3b07e9ca | 211 | schedule_delayed_work(&tg_stats_alloc_work, 0); |
ff26eaad | 212 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); |
8a3d2615 TH |
213 | } |
214 | ||
3c798398 | 215 | static void throtl_pd_exit(struct blkcg_gq *blkg) |
8a3d2615 TH |
216 | { |
217 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
ff26eaad | 218 | unsigned long flags; |
8a3d2615 | 219 | |
ff26eaad | 220 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); |
8a3d2615 | 221 | list_del_init(&tg->stats_alloc_node); |
ff26eaad | 222 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); |
8a3d2615 TH |
223 | |
224 | free_percpu(tg->stats_cpu); | |
225 | } | |
226 | ||
3c798398 | 227 | static void throtl_pd_reset_stats(struct blkcg_gq *blkg) |
8a3d2615 TH |
228 | { |
229 | struct throtl_grp *tg = blkg_to_tg(blkg); | |
230 | int cpu; | |
231 | ||
232 | if (tg->stats_cpu == NULL) | |
233 | return; | |
234 | ||
235 | for_each_possible_cpu(cpu) { | |
236 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); | |
237 | ||
238 | blkg_rwstat_reset(&sc->service_bytes); | |
239 | blkg_rwstat_reset(&sc->serviced); | |
240 | } | |
a29a171e VG |
241 | } |
242 | ||
3c798398 TH |
243 | static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td, |
244 | struct blkcg *blkcg) | |
e43473b7 | 245 | { |
be2c6b19 | 246 | /* |
3c798398 TH |
247 | * This is the common case when there are no blkcgs. Avoid lookup |
248 | * in this case | |
cd1604fa | 249 | */ |
3c798398 | 250 | if (blkcg == &blkcg_root) |
03d8e111 | 251 | return td_root_tg(td); |
e43473b7 | 252 | |
e8989fae | 253 | return blkg_to_tg(blkg_lookup(blkcg, td->queue)); |
e43473b7 VG |
254 | } |
255 | ||
cd1604fa | 256 | static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td, |
3c798398 | 257 | struct blkcg *blkcg) |
e43473b7 | 258 | { |
f469a7b4 | 259 | struct request_queue *q = td->queue; |
cd1604fa | 260 | struct throtl_grp *tg = NULL; |
bc16a4f9 | 261 | |
f469a7b4 | 262 | /* |
3c798398 TH |
263 | * This is the common case when there are no blkcgs. Avoid lookup |
264 | * in this case | |
f469a7b4 | 265 | */ |
3c798398 | 266 | if (blkcg == &blkcg_root) { |
03d8e111 | 267 | tg = td_root_tg(td); |
cd1604fa | 268 | } else { |
3c798398 | 269 | struct blkcg_gq *blkg; |
f469a7b4 | 270 | |
3c96cb32 | 271 | blkg = blkg_lookup_create(blkcg, q); |
f469a7b4 | 272 | |
cd1604fa TH |
273 | /* if %NULL and @q is alive, fall back to root_tg */ |
274 | if (!IS_ERR(blkg)) | |
0381411e | 275 | tg = blkg_to_tg(blkg); |
3f3299d5 | 276 | else if (!blk_queue_dying(q)) |
03d8e111 | 277 | tg = td_root_tg(td); |
f469a7b4 VG |
278 | } |
279 | ||
e43473b7 VG |
280 | return tg; |
281 | } | |
282 | ||
c9e0332e | 283 | static struct throtl_grp *throtl_rb_first(struct throtl_service_queue *sq) |
e43473b7 VG |
284 | { |
285 | /* Service tree is empty */ | |
c9e0332e | 286 | if (!sq->nr_pending) |
e43473b7 VG |
287 | return NULL; |
288 | ||
c9e0332e TH |
289 | if (!sq->first_pending) |
290 | sq->first_pending = rb_first(&sq->pending_tree); | |
e43473b7 | 291 | |
c9e0332e TH |
292 | if (sq->first_pending) |
293 | return rb_entry_tg(sq->first_pending); | |
e43473b7 VG |
294 | |
295 | return NULL; | |
296 | } | |
297 | ||
298 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
299 | { | |
300 | rb_erase(n, root); | |
301 | RB_CLEAR_NODE(n); | |
302 | } | |
303 | ||
c9e0332e | 304 | static void throtl_rb_erase(struct rb_node *n, struct throtl_service_queue *sq) |
e43473b7 | 305 | { |
c9e0332e TH |
306 | if (sq->first_pending == n) |
307 | sq->first_pending = NULL; | |
308 | rb_erase_init(n, &sq->pending_tree); | |
309 | --sq->nr_pending; | |
e43473b7 VG |
310 | } |
311 | ||
c9e0332e | 312 | static void update_min_dispatch_time(struct throtl_service_queue *sq) |
e43473b7 VG |
313 | { |
314 | struct throtl_grp *tg; | |
315 | ||
c9e0332e | 316 | tg = throtl_rb_first(sq); |
e43473b7 VG |
317 | if (!tg) |
318 | return; | |
319 | ||
c9e0332e | 320 | sq->first_pending_disptime = tg->disptime; |
e43473b7 VG |
321 | } |
322 | ||
c9e0332e TH |
323 | static void tg_service_queue_add(struct throtl_service_queue *sq, |
324 | struct throtl_grp *tg) | |
e43473b7 | 325 | { |
c9e0332e | 326 | struct rb_node **node = &sq->pending_tree.rb_node; |
e43473b7 VG |
327 | struct rb_node *parent = NULL; |
328 | struct throtl_grp *__tg; | |
329 | unsigned long key = tg->disptime; | |
330 | int left = 1; | |
331 | ||
332 | while (*node != NULL) { | |
333 | parent = *node; | |
334 | __tg = rb_entry_tg(parent); | |
335 | ||
336 | if (time_before(key, __tg->disptime)) | |
337 | node = &parent->rb_left; | |
338 | else { | |
339 | node = &parent->rb_right; | |
340 | left = 0; | |
341 | } | |
342 | } | |
343 | ||
344 | if (left) | |
c9e0332e | 345 | sq->first_pending = &tg->rb_node; |
e43473b7 VG |
346 | |
347 | rb_link_node(&tg->rb_node, parent, node); | |
c9e0332e | 348 | rb_insert_color(&tg->rb_node, &sq->pending_tree); |
e43473b7 VG |
349 | } |
350 | ||
351 | static void __throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
352 | { | |
c9e0332e | 353 | struct throtl_service_queue *sq = &td->service_queue; |
e43473b7 | 354 | |
c9e0332e | 355 | tg_service_queue_add(sq, tg); |
5b2c16aa | 356 | tg->flags |= THROTL_TG_PENDING; |
c9e0332e | 357 | sq->nr_pending++; |
e43473b7 VG |
358 | } |
359 | ||
360 | static void throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
361 | { | |
5b2c16aa | 362 | if (!(tg->flags & THROTL_TG_PENDING)) |
e43473b7 VG |
363 | __throtl_enqueue_tg(td, tg); |
364 | } | |
365 | ||
366 | static void __throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
367 | { | |
c9e0332e | 368 | throtl_rb_erase(&tg->rb_node, &td->service_queue); |
5b2c16aa | 369 | tg->flags &= ~THROTL_TG_PENDING; |
e43473b7 VG |
370 | } |
371 | ||
372 | static void throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg) | |
373 | { | |
5b2c16aa | 374 | if (tg->flags & THROTL_TG_PENDING) |
e43473b7 VG |
375 | __throtl_dequeue_tg(td, tg); |
376 | } | |
377 | ||
a9131a27 TH |
378 | /* Call with queue lock held */ |
379 | static void throtl_schedule_delayed_work(struct throtl_data *td, | |
380 | unsigned long delay) | |
381 | { | |
382 | struct delayed_work *dwork = &td->dispatch_work; | |
383 | ||
6a525600 TH |
384 | mod_delayed_work(kthrotld_workqueue, dwork, delay); |
385 | throtl_log(td, "schedule work. delay=%lu jiffies=%lu", delay, jiffies); | |
a9131a27 TH |
386 | } |
387 | ||
e43473b7 VG |
388 | static void throtl_schedule_next_dispatch(struct throtl_data *td) |
389 | { | |
c9e0332e | 390 | struct throtl_service_queue *sq = &td->service_queue; |
e43473b7 | 391 | |
6a525600 | 392 | /* any pending children left? */ |
c9e0332e | 393 | if (!sq->nr_pending) |
e43473b7 VG |
394 | return; |
395 | ||
c9e0332e | 396 | update_min_dispatch_time(sq); |
e43473b7 | 397 | |
c9e0332e | 398 | if (time_before_eq(sq->first_pending_disptime, jiffies)) |
450adcbe | 399 | throtl_schedule_delayed_work(td, 0); |
e43473b7 | 400 | else |
c9e0332e | 401 | throtl_schedule_delayed_work(td, sq->first_pending_disptime - jiffies); |
e43473b7 VG |
402 | } |
403 | ||
404 | static inline void | |
405 | throtl_start_new_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
406 | { | |
407 | tg->bytes_disp[rw] = 0; | |
8e89d13f | 408 | tg->io_disp[rw] = 0; |
e43473b7 VG |
409 | tg->slice_start[rw] = jiffies; |
410 | tg->slice_end[rw] = jiffies + throtl_slice; | |
411 | throtl_log_tg(td, tg, "[%c] new slice start=%lu end=%lu jiffies=%lu", | |
412 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
413 | tg->slice_end[rw], jiffies); | |
414 | } | |
415 | ||
d1ae8ffd VG |
416 | static inline void throtl_set_slice_end(struct throtl_data *td, |
417 | struct throtl_grp *tg, bool rw, unsigned long jiffy_end) | |
418 | { | |
419 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
420 | } | |
421 | ||
e43473b7 VG |
422 | static inline void throtl_extend_slice(struct throtl_data *td, |
423 | struct throtl_grp *tg, bool rw, unsigned long jiffy_end) | |
424 | { | |
425 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); | |
426 | throtl_log_tg(td, tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu", | |
427 | rw == READ ? 'R' : 'W', tg->slice_start[rw], | |
428 | tg->slice_end[rw], jiffies); | |
429 | } | |
430 | ||
431 | /* Determine if previously allocated or extended slice is complete or not */ | |
432 | static bool | |
433 | throtl_slice_used(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
434 | { | |
435 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) | |
436 | return 0; | |
437 | ||
438 | return 1; | |
439 | } | |
440 | ||
441 | /* Trim the used slices and adjust slice start accordingly */ | |
442 | static inline void | |
443 | throtl_trim_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw) | |
444 | { | |
3aad5d3e VG |
445 | unsigned long nr_slices, time_elapsed, io_trim; |
446 | u64 bytes_trim, tmp; | |
e43473b7 VG |
447 | |
448 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); | |
449 | ||
450 | /* | |
451 | * If bps are unlimited (-1), then time slice don't get | |
452 | * renewed. Don't try to trim the slice if slice is used. A new | |
453 | * slice will start when appropriate. | |
454 | */ | |
455 | if (throtl_slice_used(td, tg, rw)) | |
456 | return; | |
457 | ||
d1ae8ffd VG |
458 | /* |
459 | * A bio has been dispatched. Also adjust slice_end. It might happen | |
460 | * that initially cgroup limit was very low resulting in high | |
461 | * slice_end, but later limit was bumped up and bio was dispached | |
462 | * sooner, then we need to reduce slice_end. A high bogus slice_end | |
463 | * is bad because it does not allow new slice to start. | |
464 | */ | |
465 | ||
466 | throtl_set_slice_end(td, tg, rw, jiffies + throtl_slice); | |
467 | ||
e43473b7 VG |
468 | time_elapsed = jiffies - tg->slice_start[rw]; |
469 | ||
470 | nr_slices = time_elapsed / throtl_slice; | |
471 | ||
472 | if (!nr_slices) | |
473 | return; | |
3aad5d3e VG |
474 | tmp = tg->bps[rw] * throtl_slice * nr_slices; |
475 | do_div(tmp, HZ); | |
476 | bytes_trim = tmp; | |
e43473b7 | 477 | |
8e89d13f | 478 | io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ; |
e43473b7 | 479 | |
8e89d13f | 480 | if (!bytes_trim && !io_trim) |
e43473b7 VG |
481 | return; |
482 | ||
483 | if (tg->bytes_disp[rw] >= bytes_trim) | |
484 | tg->bytes_disp[rw] -= bytes_trim; | |
485 | else | |
486 | tg->bytes_disp[rw] = 0; | |
487 | ||
8e89d13f VG |
488 | if (tg->io_disp[rw] >= io_trim) |
489 | tg->io_disp[rw] -= io_trim; | |
490 | else | |
491 | tg->io_disp[rw] = 0; | |
492 | ||
e43473b7 VG |
493 | tg->slice_start[rw] += nr_slices * throtl_slice; |
494 | ||
3aad5d3e | 495 | throtl_log_tg(td, tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu" |
e43473b7 | 496 | " start=%lu end=%lu jiffies=%lu", |
8e89d13f | 497 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, |
e43473b7 VG |
498 | tg->slice_start[rw], tg->slice_end[rw], jiffies); |
499 | } | |
500 | ||
8e89d13f VG |
501 | static bool tg_with_in_iops_limit(struct throtl_data *td, struct throtl_grp *tg, |
502 | struct bio *bio, unsigned long *wait) | |
e43473b7 VG |
503 | { |
504 | bool rw = bio_data_dir(bio); | |
8e89d13f | 505 | unsigned int io_allowed; |
e43473b7 | 506 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
c49c06e4 | 507 | u64 tmp; |
e43473b7 | 508 | |
8e89d13f | 509 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
e43473b7 | 510 | |
8e89d13f VG |
511 | /* Slice has just started. Consider one slice interval */ |
512 | if (!jiffy_elapsed) | |
513 | jiffy_elapsed_rnd = throtl_slice; | |
514 | ||
515 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
516 | ||
c49c06e4 VG |
517 | /* |
518 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be | |
519 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we | |
520 | * will allow dispatch after 1 second and after that slice should | |
521 | * have been trimmed. | |
522 | */ | |
523 | ||
524 | tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd; | |
525 | do_div(tmp, HZ); | |
526 | ||
527 | if (tmp > UINT_MAX) | |
528 | io_allowed = UINT_MAX; | |
529 | else | |
530 | io_allowed = tmp; | |
8e89d13f VG |
531 | |
532 | if (tg->io_disp[rw] + 1 <= io_allowed) { | |
e43473b7 VG |
533 | if (wait) |
534 | *wait = 0; | |
535 | return 1; | |
536 | } | |
537 | ||
8e89d13f VG |
538 | /* Calc approx time to dispatch */ |
539 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1; | |
540 | ||
541 | if (jiffy_wait > jiffy_elapsed) | |
542 | jiffy_wait = jiffy_wait - jiffy_elapsed; | |
543 | else | |
544 | jiffy_wait = 1; | |
545 | ||
546 | if (wait) | |
547 | *wait = jiffy_wait; | |
548 | return 0; | |
549 | } | |
550 | ||
551 | static bool tg_with_in_bps_limit(struct throtl_data *td, struct throtl_grp *tg, | |
552 | struct bio *bio, unsigned long *wait) | |
553 | { | |
554 | bool rw = bio_data_dir(bio); | |
3aad5d3e | 555 | u64 bytes_allowed, extra_bytes, tmp; |
8e89d13f | 556 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
e43473b7 VG |
557 | |
558 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; | |
559 | ||
560 | /* Slice has just started. Consider one slice interval */ | |
561 | if (!jiffy_elapsed) | |
562 | jiffy_elapsed_rnd = throtl_slice; | |
563 | ||
564 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); | |
565 | ||
5e901a2b VG |
566 | tmp = tg->bps[rw] * jiffy_elapsed_rnd; |
567 | do_div(tmp, HZ); | |
3aad5d3e | 568 | bytes_allowed = tmp; |
e43473b7 VG |
569 | |
570 | if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) { | |
571 | if (wait) | |
572 | *wait = 0; | |
573 | return 1; | |
574 | } | |
575 | ||
576 | /* Calc approx time to dispatch */ | |
577 | extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed; | |
578 | jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]); | |
579 | ||
580 | if (!jiffy_wait) | |
581 | jiffy_wait = 1; | |
582 | ||
583 | /* | |
584 | * This wait time is without taking into consideration the rounding | |
585 | * up we did. Add that time also. | |
586 | */ | |
587 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); | |
e43473b7 VG |
588 | if (wait) |
589 | *wait = jiffy_wait; | |
8e89d13f VG |
590 | return 0; |
591 | } | |
592 | ||
af75cd3c VG |
593 | static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) { |
594 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) | |
595 | return 1; | |
596 | return 0; | |
597 | } | |
598 | ||
8e89d13f VG |
599 | /* |
600 | * Returns whether one can dispatch a bio or not. Also returns approx number | |
601 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched | |
602 | */ | |
603 | static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg, | |
604 | struct bio *bio, unsigned long *wait) | |
605 | { | |
606 | bool rw = bio_data_dir(bio); | |
607 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; | |
608 | ||
609 | /* | |
610 | * Currently whole state machine of group depends on first bio | |
611 | * queued in the group bio list. So one should not be calling | |
612 | * this function with a different bio if there are other bios | |
613 | * queued. | |
614 | */ | |
615 | BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw])); | |
e43473b7 | 616 | |
8e89d13f VG |
617 | /* If tg->bps = -1, then BW is unlimited */ |
618 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) { | |
619 | if (wait) | |
620 | *wait = 0; | |
621 | return 1; | |
622 | } | |
623 | ||
624 | /* | |
625 | * If previous slice expired, start a new one otherwise renew/extend | |
626 | * existing slice to make sure it is at least throtl_slice interval | |
627 | * long since now. | |
628 | */ | |
629 | if (throtl_slice_used(td, tg, rw)) | |
630 | throtl_start_new_slice(td, tg, rw); | |
631 | else { | |
632 | if (time_before(tg->slice_end[rw], jiffies + throtl_slice)) | |
633 | throtl_extend_slice(td, tg, rw, jiffies + throtl_slice); | |
634 | } | |
635 | ||
636 | if (tg_with_in_bps_limit(td, tg, bio, &bps_wait) | |
637 | && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) { | |
638 | if (wait) | |
639 | *wait = 0; | |
640 | return 1; | |
641 | } | |
642 | ||
643 | max_wait = max(bps_wait, iops_wait); | |
644 | ||
645 | if (wait) | |
646 | *wait = max_wait; | |
647 | ||
648 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) | |
649 | throtl_extend_slice(td, tg, rw, jiffies + max_wait); | |
e43473b7 VG |
650 | |
651 | return 0; | |
652 | } | |
653 | ||
3c798398 | 654 | static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes, |
629ed0b1 TH |
655 | int rw) |
656 | { | |
8a3d2615 TH |
657 | struct throtl_grp *tg = blkg_to_tg(blkg); |
658 | struct tg_stats_cpu *stats_cpu; | |
629ed0b1 TH |
659 | unsigned long flags; |
660 | ||
661 | /* If per cpu stats are not allocated yet, don't do any accounting. */ | |
8a3d2615 | 662 | if (tg->stats_cpu == NULL) |
629ed0b1 TH |
663 | return; |
664 | ||
665 | /* | |
666 | * Disabling interrupts to provide mutual exclusion between two | |
667 | * writes on same cpu. It probably is not needed for 64bit. Not | |
668 | * optimizing that case yet. | |
669 | */ | |
670 | local_irq_save(flags); | |
671 | ||
8a3d2615 | 672 | stats_cpu = this_cpu_ptr(tg->stats_cpu); |
629ed0b1 | 673 | |
629ed0b1 TH |
674 | blkg_rwstat_add(&stats_cpu->serviced, rw, 1); |
675 | blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes); | |
676 | ||
677 | local_irq_restore(flags); | |
678 | } | |
679 | ||
e43473b7 VG |
680 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) |
681 | { | |
682 | bool rw = bio_data_dir(bio); | |
e43473b7 VG |
683 | |
684 | /* Charge the bio to the group */ | |
685 | tg->bytes_disp[rw] += bio->bi_size; | |
8e89d13f | 686 | tg->io_disp[rw]++; |
e43473b7 | 687 | |
629ed0b1 | 688 | throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size, bio->bi_rw); |
e43473b7 VG |
689 | } |
690 | ||
691 | static void throtl_add_bio_tg(struct throtl_data *td, struct throtl_grp *tg, | |
692 | struct bio *bio) | |
693 | { | |
694 | bool rw = bio_data_dir(bio); | |
695 | ||
696 | bio_list_add(&tg->bio_lists[rw], bio); | |
697 | /* Take a bio reference on tg */ | |
1adaf3dd | 698 | blkg_get(tg_to_blkg(tg)); |
e43473b7 VG |
699 | tg->nr_queued[rw]++; |
700 | td->nr_queued[rw]++; | |
701 | throtl_enqueue_tg(td, tg); | |
702 | } | |
703 | ||
704 | static void tg_update_disptime(struct throtl_data *td, struct throtl_grp *tg) | |
705 | { | |
706 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; | |
707 | struct bio *bio; | |
708 | ||
709 | if ((bio = bio_list_peek(&tg->bio_lists[READ]))) | |
710 | tg_may_dispatch(td, tg, bio, &read_wait); | |
711 | ||
712 | if ((bio = bio_list_peek(&tg->bio_lists[WRITE]))) | |
713 | tg_may_dispatch(td, tg, bio, &write_wait); | |
714 | ||
715 | min_wait = min(read_wait, write_wait); | |
716 | disptime = jiffies + min_wait; | |
717 | ||
e43473b7 VG |
718 | /* Update dispatch time */ |
719 | throtl_dequeue_tg(td, tg); | |
720 | tg->disptime = disptime; | |
721 | throtl_enqueue_tg(td, tg); | |
722 | } | |
723 | ||
724 | static void tg_dispatch_one_bio(struct throtl_data *td, struct throtl_grp *tg, | |
725 | bool rw, struct bio_list *bl) | |
726 | { | |
727 | struct bio *bio; | |
728 | ||
729 | bio = bio_list_pop(&tg->bio_lists[rw]); | |
730 | tg->nr_queued[rw]--; | |
1adaf3dd TH |
731 | /* Drop bio reference on blkg */ |
732 | blkg_put(tg_to_blkg(tg)); | |
e43473b7 VG |
733 | |
734 | BUG_ON(td->nr_queued[rw] <= 0); | |
735 | td->nr_queued[rw]--; | |
736 | ||
737 | throtl_charge_bio(tg, bio); | |
738 | bio_list_add(bl, bio); | |
739 | bio->bi_rw |= REQ_THROTTLED; | |
740 | ||
741 | throtl_trim_slice(td, tg, rw); | |
742 | } | |
743 | ||
744 | static int throtl_dispatch_tg(struct throtl_data *td, struct throtl_grp *tg, | |
745 | struct bio_list *bl) | |
746 | { | |
747 | unsigned int nr_reads = 0, nr_writes = 0; | |
748 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; | |
c2f6805d | 749 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; |
e43473b7 VG |
750 | struct bio *bio; |
751 | ||
752 | /* Try to dispatch 75% READS and 25% WRITES */ | |
753 | ||
754 | while ((bio = bio_list_peek(&tg->bio_lists[READ])) | |
755 | && tg_may_dispatch(td, tg, bio, NULL)) { | |
756 | ||
757 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl); | |
758 | nr_reads++; | |
759 | ||
760 | if (nr_reads >= max_nr_reads) | |
761 | break; | |
762 | } | |
763 | ||
764 | while ((bio = bio_list_peek(&tg->bio_lists[WRITE])) | |
765 | && tg_may_dispatch(td, tg, bio, NULL)) { | |
766 | ||
767 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl); | |
768 | nr_writes++; | |
769 | ||
770 | if (nr_writes >= max_nr_writes) | |
771 | break; | |
772 | } | |
773 | ||
774 | return nr_reads + nr_writes; | |
775 | } | |
776 | ||
777 | static int throtl_select_dispatch(struct throtl_data *td, struct bio_list *bl) | |
778 | { | |
779 | unsigned int nr_disp = 0; | |
780 | struct throtl_grp *tg; | |
c9e0332e | 781 | struct throtl_service_queue *sq = &td->service_queue; |
e43473b7 VG |
782 | |
783 | while (1) { | |
c9e0332e | 784 | tg = throtl_rb_first(sq); |
e43473b7 VG |
785 | |
786 | if (!tg) | |
787 | break; | |
788 | ||
789 | if (time_before(jiffies, tg->disptime)) | |
790 | break; | |
791 | ||
792 | throtl_dequeue_tg(td, tg); | |
793 | ||
794 | nr_disp += throtl_dispatch_tg(td, tg, bl); | |
795 | ||
2db6314c | 796 | if (tg->nr_queued[0] || tg->nr_queued[1]) |
e43473b7 | 797 | tg_update_disptime(td, tg); |
e43473b7 VG |
798 | |
799 | if (nr_disp >= throtl_quantum) | |
800 | break; | |
801 | } | |
802 | ||
803 | return nr_disp; | |
804 | } | |
805 | ||
cb76199c TH |
806 | /* work function to dispatch throttled bios */ |
807 | void blk_throtl_dispatch_work_fn(struct work_struct *work) | |
e43473b7 | 808 | { |
cb76199c TH |
809 | struct throtl_data *td = container_of(to_delayed_work(work), |
810 | struct throtl_data, dispatch_work); | |
811 | struct request_queue *q = td->queue; | |
e43473b7 VG |
812 | unsigned int nr_disp = 0; |
813 | struct bio_list bio_list_on_stack; | |
814 | struct bio *bio; | |
69d60eb9 | 815 | struct blk_plug plug; |
e43473b7 VG |
816 | |
817 | spin_lock_irq(q->queue_lock); | |
818 | ||
e43473b7 VG |
819 | bio_list_init(&bio_list_on_stack); |
820 | ||
d2f31a5f | 821 | throtl_log(td, "dispatch nr_queued=%u read=%u write=%u", |
6a525600 TH |
822 | td->nr_queued[READ] + td->nr_queued[WRITE], |
823 | td->nr_queued[READ], td->nr_queued[WRITE]); | |
e43473b7 VG |
824 | |
825 | nr_disp = throtl_select_dispatch(td, &bio_list_on_stack); | |
826 | ||
827 | if (nr_disp) | |
828 | throtl_log(td, "bios disp=%u", nr_disp); | |
829 | ||
830 | throtl_schedule_next_dispatch(td); | |
6a525600 | 831 | |
e43473b7 VG |
832 | spin_unlock_irq(q->queue_lock); |
833 | ||
834 | /* | |
835 | * If we dispatched some requests, unplug the queue to make sure | |
836 | * immediate dispatch | |
837 | */ | |
838 | if (nr_disp) { | |
69d60eb9 | 839 | blk_start_plug(&plug); |
e43473b7 VG |
840 | while((bio = bio_list_pop(&bio_list_on_stack))) |
841 | generic_make_request(bio); | |
69d60eb9 | 842 | blk_finish_plug(&plug); |
e43473b7 | 843 | } |
e43473b7 VG |
844 | } |
845 | ||
f95a04af TH |
846 | static u64 tg_prfill_cpu_rwstat(struct seq_file *sf, |
847 | struct blkg_policy_data *pd, int off) | |
41b38b6d | 848 | { |
f95a04af | 849 | struct throtl_grp *tg = pd_to_tg(pd); |
41b38b6d TH |
850 | struct blkg_rwstat rwstat = { }, tmp; |
851 | int i, cpu; | |
852 | ||
853 | for_each_possible_cpu(cpu) { | |
8a3d2615 | 854 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); |
41b38b6d TH |
855 | |
856 | tmp = blkg_rwstat_read((void *)sc + off); | |
857 | for (i = 0; i < BLKG_RWSTAT_NR; i++) | |
858 | rwstat.cnt[i] += tmp.cnt[i]; | |
859 | } | |
860 | ||
f95a04af | 861 | return __blkg_prfill_rwstat(sf, pd, &rwstat); |
41b38b6d TH |
862 | } |
863 | ||
8a3d2615 TH |
864 | static int tg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft, |
865 | struct seq_file *sf) | |
41b38b6d | 866 | { |
3c798398 | 867 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); |
41b38b6d | 868 | |
3c798398 | 869 | blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkcg_policy_throtl, |
5bc4afb1 | 870 | cft->private, true); |
41b38b6d TH |
871 | return 0; |
872 | } | |
873 | ||
f95a04af TH |
874 | static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd, |
875 | int off) | |
60c2bc2d | 876 | { |
f95a04af TH |
877 | struct throtl_grp *tg = pd_to_tg(pd); |
878 | u64 v = *(u64 *)((void *)tg + off); | |
60c2bc2d | 879 | |
af133ceb | 880 | if (v == -1) |
60c2bc2d | 881 | return 0; |
f95a04af | 882 | return __blkg_prfill_u64(sf, pd, v); |
60c2bc2d TH |
883 | } |
884 | ||
f95a04af TH |
885 | static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd, |
886 | int off) | |
e43473b7 | 887 | { |
f95a04af TH |
888 | struct throtl_grp *tg = pd_to_tg(pd); |
889 | unsigned int v = *(unsigned int *)((void *)tg + off); | |
fe071437 | 890 | |
af133ceb TH |
891 | if (v == -1) |
892 | return 0; | |
f95a04af | 893 | return __blkg_prfill_u64(sf, pd, v); |
e43473b7 VG |
894 | } |
895 | ||
af133ceb TH |
896 | static int tg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
897 | struct seq_file *sf) | |
8e89d13f | 898 | { |
3c798398 TH |
899 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_u64, |
900 | &blkcg_policy_throtl, cft->private, false); | |
af133ceb | 901 | return 0; |
8e89d13f VG |
902 | } |
903 | ||
af133ceb TH |
904 | static int tg_print_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
905 | struct seq_file *sf) | |
8e89d13f | 906 | { |
3c798398 TH |
907 | blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_uint, |
908 | &blkcg_policy_throtl, cft->private, false); | |
af133ceb | 909 | return 0; |
60c2bc2d TH |
910 | } |
911 | ||
af133ceb TH |
912 | static int tg_set_conf(struct cgroup *cgrp, struct cftype *cft, const char *buf, |
913 | bool is_u64) | |
60c2bc2d | 914 | { |
3c798398 | 915 | struct blkcg *blkcg = cgroup_to_blkcg(cgrp); |
60c2bc2d | 916 | struct blkg_conf_ctx ctx; |
af133ceb | 917 | struct throtl_grp *tg; |
a2b1693b | 918 | struct throtl_data *td; |
60c2bc2d TH |
919 | int ret; |
920 | ||
3c798398 | 921 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); |
60c2bc2d TH |
922 | if (ret) |
923 | return ret; | |
924 | ||
af133ceb | 925 | tg = blkg_to_tg(ctx.blkg); |
a2b1693b | 926 | td = ctx.blkg->q->td; |
af133ceb | 927 | |
a2b1693b TH |
928 | if (!ctx.v) |
929 | ctx.v = -1; | |
af133ceb | 930 | |
a2b1693b TH |
931 | if (is_u64) |
932 | *(u64 *)((void *)tg + cft->private) = ctx.v; | |
933 | else | |
934 | *(unsigned int *)((void *)tg + cft->private) = ctx.v; | |
af133ceb | 935 | |
632b4493 TH |
936 | throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu riops=%u wiops=%u", |
937 | tg->bps[READ], tg->bps[WRITE], | |
938 | tg->iops[READ], tg->iops[WRITE]); | |
939 | ||
940 | /* | |
941 | * We're already holding queue_lock and know @tg is valid. Let's | |
942 | * apply the new config directly. | |
943 | * | |
944 | * Restart the slices for both READ and WRITES. It might happen | |
945 | * that a group's limit are dropped suddenly and we don't want to | |
946 | * account recently dispatched IO with new low rate. | |
947 | */ | |
948 | throtl_start_new_slice(td, tg, 0); | |
949 | throtl_start_new_slice(td, tg, 1); | |
950 | ||
5b2c16aa | 951 | if (tg->flags & THROTL_TG_PENDING) { |
632b4493 TH |
952 | tg_update_disptime(td, tg); |
953 | throtl_schedule_next_dispatch(td); | |
954 | } | |
60c2bc2d TH |
955 | |
956 | blkg_conf_finish(&ctx); | |
a2b1693b | 957 | return 0; |
8e89d13f VG |
958 | } |
959 | ||
af133ceb TH |
960 | static int tg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft, |
961 | const char *buf) | |
60c2bc2d | 962 | { |
af133ceb | 963 | return tg_set_conf(cgrp, cft, buf, true); |
60c2bc2d TH |
964 | } |
965 | ||
af133ceb TH |
966 | static int tg_set_conf_uint(struct cgroup *cgrp, struct cftype *cft, |
967 | const char *buf) | |
60c2bc2d | 968 | { |
af133ceb | 969 | return tg_set_conf(cgrp, cft, buf, false); |
60c2bc2d TH |
970 | } |
971 | ||
972 | static struct cftype throtl_files[] = { | |
973 | { | |
974 | .name = "throttle.read_bps_device", | |
af133ceb TH |
975 | .private = offsetof(struct throtl_grp, bps[READ]), |
976 | .read_seq_string = tg_print_conf_u64, | |
977 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
978 | .max_write_len = 256, |
979 | }, | |
980 | { | |
981 | .name = "throttle.write_bps_device", | |
af133ceb TH |
982 | .private = offsetof(struct throtl_grp, bps[WRITE]), |
983 | .read_seq_string = tg_print_conf_u64, | |
984 | .write_string = tg_set_conf_u64, | |
60c2bc2d TH |
985 | .max_write_len = 256, |
986 | }, | |
987 | { | |
988 | .name = "throttle.read_iops_device", | |
af133ceb TH |
989 | .private = offsetof(struct throtl_grp, iops[READ]), |
990 | .read_seq_string = tg_print_conf_uint, | |
991 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
992 | .max_write_len = 256, |
993 | }, | |
994 | { | |
995 | .name = "throttle.write_iops_device", | |
af133ceb TH |
996 | .private = offsetof(struct throtl_grp, iops[WRITE]), |
997 | .read_seq_string = tg_print_conf_uint, | |
998 | .write_string = tg_set_conf_uint, | |
60c2bc2d TH |
999 | .max_write_len = 256, |
1000 | }, | |
1001 | { | |
1002 | .name = "throttle.io_service_bytes", | |
5bc4afb1 | 1003 | .private = offsetof(struct tg_stats_cpu, service_bytes), |
8a3d2615 | 1004 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1005 | }, |
1006 | { | |
1007 | .name = "throttle.io_serviced", | |
5bc4afb1 | 1008 | .private = offsetof(struct tg_stats_cpu, serviced), |
8a3d2615 | 1009 | .read_seq_string = tg_print_cpu_rwstat, |
60c2bc2d TH |
1010 | }, |
1011 | { } /* terminate */ | |
1012 | }; | |
1013 | ||
da527770 | 1014 | static void throtl_shutdown_wq(struct request_queue *q) |
e43473b7 VG |
1015 | { |
1016 | struct throtl_data *td = q->td; | |
1017 | ||
cb76199c | 1018 | cancel_delayed_work_sync(&td->dispatch_work); |
e43473b7 VG |
1019 | } |
1020 | ||
3c798398 | 1021 | static struct blkcg_policy blkcg_policy_throtl = { |
f9fcc2d3 TH |
1022 | .pd_size = sizeof(struct throtl_grp), |
1023 | .cftypes = throtl_files, | |
1024 | ||
1025 | .pd_init_fn = throtl_pd_init, | |
1026 | .pd_exit_fn = throtl_pd_exit, | |
1027 | .pd_reset_stats_fn = throtl_pd_reset_stats, | |
e43473b7 VG |
1028 | }; |
1029 | ||
bc16a4f9 | 1030 | bool blk_throtl_bio(struct request_queue *q, struct bio *bio) |
e43473b7 VG |
1031 | { |
1032 | struct throtl_data *td = q->td; | |
1033 | struct throtl_grp *tg; | |
e43473b7 | 1034 | bool rw = bio_data_dir(bio), update_disptime = true; |
3c798398 | 1035 | struct blkcg *blkcg; |
bc16a4f9 | 1036 | bool throttled = false; |
e43473b7 VG |
1037 | |
1038 | if (bio->bi_rw & REQ_THROTTLED) { | |
1039 | bio->bi_rw &= ~REQ_THROTTLED; | |
bc16a4f9 | 1040 | goto out; |
e43473b7 VG |
1041 | } |
1042 | ||
af75cd3c VG |
1043 | /* |
1044 | * A throtl_grp pointer retrieved under rcu can be used to access | |
1045 | * basic fields like stats and io rates. If a group has no rules, | |
1046 | * just update the dispatch stats in lockless manner and return. | |
1047 | */ | |
af75cd3c | 1048 | rcu_read_lock(); |
3c798398 | 1049 | blkcg = bio_blkcg(bio); |
cd1604fa | 1050 | tg = throtl_lookup_tg(td, blkcg); |
af75cd3c | 1051 | if (tg) { |
af75cd3c | 1052 | if (tg_no_rule_group(tg, rw)) { |
629ed0b1 TH |
1053 | throtl_update_dispatch_stats(tg_to_blkg(tg), |
1054 | bio->bi_size, bio->bi_rw); | |
2a7f1244 | 1055 | goto out_unlock_rcu; |
af75cd3c VG |
1056 | } |
1057 | } | |
af75cd3c VG |
1058 | |
1059 | /* | |
1060 | * Either group has not been allocated yet or it is not an unlimited | |
1061 | * IO group | |
1062 | */ | |
e43473b7 | 1063 | spin_lock_irq(q->queue_lock); |
cd1604fa | 1064 | tg = throtl_lookup_create_tg(td, blkcg); |
bc16a4f9 TH |
1065 | if (unlikely(!tg)) |
1066 | goto out_unlock; | |
f469a7b4 | 1067 | |
e43473b7 VG |
1068 | if (tg->nr_queued[rw]) { |
1069 | /* | |
1070 | * There is already another bio queued in same dir. No | |
1071 | * need to update dispatch time. | |
1072 | */ | |
231d704b | 1073 | update_disptime = false; |
e43473b7 | 1074 | goto queue_bio; |
de701c74 | 1075 | |
e43473b7 VG |
1076 | } |
1077 | ||
1078 | /* Bio is with-in rate limit of group */ | |
1079 | if (tg_may_dispatch(td, tg, bio, NULL)) { | |
1080 | throtl_charge_bio(tg, bio); | |
04521db0 VG |
1081 | |
1082 | /* | |
1083 | * We need to trim slice even when bios are not being queued | |
1084 | * otherwise it might happen that a bio is not queued for | |
1085 | * a long time and slice keeps on extending and trim is not | |
1086 | * called for a long time. Now if limits are reduced suddenly | |
1087 | * we take into account all the IO dispatched so far at new | |
1088 | * low rate and * newly queued IO gets a really long dispatch | |
1089 | * time. | |
1090 | * | |
1091 | * So keep on trimming slice even if bio is not queued. | |
1092 | */ | |
1093 | throtl_trim_slice(td, tg, rw); | |
bc16a4f9 | 1094 | goto out_unlock; |
e43473b7 VG |
1095 | } |
1096 | ||
1097 | queue_bio: | |
fd16d263 | 1098 | throtl_log_tg(td, tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu" |
8e89d13f VG |
1099 | " iodisp=%u iops=%u queued=%d/%d", |
1100 | rw == READ ? 'R' : 'W', | |
e43473b7 | 1101 | tg->bytes_disp[rw], bio->bi_size, tg->bps[rw], |
8e89d13f | 1102 | tg->io_disp[rw], tg->iops[rw], |
e43473b7 VG |
1103 | tg->nr_queued[READ], tg->nr_queued[WRITE]); |
1104 | ||
671058fb | 1105 | bio_associate_current(bio); |
e43473b7 | 1106 | throtl_add_bio_tg(q->td, tg, bio); |
bc16a4f9 | 1107 | throttled = true; |
e43473b7 VG |
1108 | |
1109 | if (update_disptime) { | |
1110 | tg_update_disptime(td, tg); | |
1111 | throtl_schedule_next_dispatch(td); | |
1112 | } | |
1113 | ||
bc16a4f9 | 1114 | out_unlock: |
e43473b7 | 1115 | spin_unlock_irq(q->queue_lock); |
2a7f1244 TH |
1116 | out_unlock_rcu: |
1117 | rcu_read_unlock(); | |
bc16a4f9 TH |
1118 | out: |
1119 | return throttled; | |
e43473b7 VG |
1120 | } |
1121 | ||
c9a929dd TH |
1122 | /** |
1123 | * blk_throtl_drain - drain throttled bios | |
1124 | * @q: request_queue to drain throttled bios for | |
1125 | * | |
1126 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). | |
1127 | */ | |
1128 | void blk_throtl_drain(struct request_queue *q) | |
1129 | __releases(q->queue_lock) __acquires(q->queue_lock) | |
1130 | { | |
1131 | struct throtl_data *td = q->td; | |
c9e0332e | 1132 | struct throtl_service_queue *sq = &td->service_queue; |
c9a929dd TH |
1133 | struct throtl_grp *tg; |
1134 | struct bio_list bl; | |
1135 | struct bio *bio; | |
1136 | ||
8bcb6c7d | 1137 | queue_lockdep_assert_held(q); |
c9a929dd TH |
1138 | |
1139 | bio_list_init(&bl); | |
1140 | ||
c9e0332e | 1141 | while ((tg = throtl_rb_first(sq))) { |
c9a929dd TH |
1142 | throtl_dequeue_tg(td, tg); |
1143 | ||
1144 | while ((bio = bio_list_peek(&tg->bio_lists[READ]))) | |
1145 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl); | |
1146 | while ((bio = bio_list_peek(&tg->bio_lists[WRITE]))) | |
1147 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl); | |
1148 | } | |
1149 | spin_unlock_irq(q->queue_lock); | |
1150 | ||
1151 | while ((bio = bio_list_pop(&bl))) | |
1152 | generic_make_request(bio); | |
1153 | ||
1154 | spin_lock_irq(q->queue_lock); | |
1155 | } | |
1156 | ||
e43473b7 VG |
1157 | int blk_throtl_init(struct request_queue *q) |
1158 | { | |
1159 | struct throtl_data *td; | |
a2b1693b | 1160 | int ret; |
e43473b7 VG |
1161 | |
1162 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); | |
1163 | if (!td) | |
1164 | return -ENOMEM; | |
1165 | ||
c9e0332e | 1166 | td->service_queue = THROTL_SERVICE_QUEUE_INITIALIZER; |
cb76199c | 1167 | INIT_DELAYED_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn); |
e43473b7 | 1168 | |
cd1604fa | 1169 | q->td = td; |
29b12589 | 1170 | td->queue = q; |
02977e4a | 1171 | |
a2b1693b | 1172 | /* activate policy */ |
3c798398 | 1173 | ret = blkcg_activate_policy(q, &blkcg_policy_throtl); |
a2b1693b | 1174 | if (ret) |
f51b802c | 1175 | kfree(td); |
a2b1693b | 1176 | return ret; |
e43473b7 VG |
1177 | } |
1178 | ||
1179 | void blk_throtl_exit(struct request_queue *q) | |
1180 | { | |
c875f4d0 | 1181 | BUG_ON(!q->td); |
da527770 | 1182 | throtl_shutdown_wq(q); |
3c798398 | 1183 | blkcg_deactivate_policy(q, &blkcg_policy_throtl); |
c9a929dd | 1184 | kfree(q->td); |
e43473b7 VG |
1185 | } |
1186 | ||
1187 | static int __init throtl_init(void) | |
1188 | { | |
450adcbe VG |
1189 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
1190 | if (!kthrotld_workqueue) | |
1191 | panic("Failed to create kthrotld\n"); | |
1192 | ||
3c798398 | 1193 | return blkcg_policy_register(&blkcg_policy_throtl); |
e43473b7 VG |
1194 | } |
1195 | ||
1196 | module_init(throtl_init); |